Consequently, we suggest the policy considering, initially, the use of a 15%% relative cutoff point, followed by a restriction accepting only terms with a relevance index above or equal to 2, i.e., a composition with a threshold-based cutoff point.
As seen before, a 15%% relative cutoff point would be a good trade-off for bigrams and trigrams of the pediatrics corpus.
That being said, there are three traditional approaches to guess a cutoff point: absolute, threshold-based, and relative cutoff points.
The specifics of this filter depend on knowledge of the data at hand, e.g. removing genes never significantly above background in a microarray dataset or metagenomic taxa never above a relative abundance cutoff.
It is important to notice that unlike the other traditional approaches, relative cutoff points seem to deliver a larger reasonable balanced region.
As in the other approaches, it is still necessary to arbitrarily choose the percentage of terms to consider, but the biggest advantage of the relative cutoff approach is to offer an option that is independent of the extraction process or corpora size, even thought it is still dependent on the size of the extracted term list.
Consequently, it seems that even relative cutoff points cannot be considered alone as an effective way to estimate a precision-recall balanced point for all extracted term lists, despite the finding of reasonable values of F-measure (above 75%%) for values of 14 and 15 % in bigrams and trigrams lists.
A rarely found alternative [17] to the absolute and threshold approaches is the use of relative cutoff points.
Figure 4 depicts the precision, recall, and F-measure values obtained for these relative cutoff points (numeric values at Table 6).
